Designing Software in the Large(dafoster.net)
dafoster.net
Designing Software in the Large
https://dafoster.net/articles/2025/07/22/designing-software-in-the-large/
40 comments
> Just taking the notion of "complexity". Reducing that to _just_ cognitive load seems to be a very poor analysis, when simple/complex ought to deal with the "size" of a structure, not how easy it is to understand.
Preface: I'm likely nitpicking here; the use of "_just_" is enough for me to mostly agree with your take.
Isn't the idea that the bulk of complexity IS in the understanding of how a system works, both how it should work and how it does work? We could take the Quake Fast Inverse Square Root code, which is simple in "size" but quite complex on how it actually achieves its outcome. I'd argue it requires comments, tests, and/or clarifications to make sense of what its actually doing.
How do we measure that complexity? No idea :) But I like to believe that's why the book takes a philosophical approach to the discussion.
I agree the arguments in the book largely "make sense" to me but I found myself finding it a little hand-wavey on it actually proving its points without concrete examples. I don't recall there being any metrics or measurements on improvement either, making it a philosophical discussion to me and not a scientific exercise.
Preface: I'm likely nitpicking here; the use of "_just_" is enough for me to mostly agree with your take.
Isn't the idea that the bulk of complexity IS in the understanding of how a system works, both how it should work and how it does work? We could take the Quake Fast Inverse Square Root code, which is simple in "size" but quite complex on how it actually achieves its outcome. I'd argue it requires comments, tests, and/or clarifications to make sense of what its actually doing.
How do we measure that complexity? No idea :) But I like to believe that's why the book takes a philosophical approach to the discussion.
I agree the arguments in the book largely "make sense" to me but I found myself finding it a little hand-wavey on it actually proving its points without concrete examples. I don't recall there being any metrics or measurements on improvement either, making it a philosophical discussion to me and not a scientific exercise.
I mean, we can definitively talk about simplicity/complexity in a fairly easy way when it comes to mathematical structures or data structures in my opinion.
For instance, a binary tree that contains just a root node is clearly simpler than a binary tree with three nodes, if we take "simple" to mean "with less parts" and complex to mean "with more parts". Similarly, a "molecule" is more complex than an "atom".
This is a useful definition, I think, because when we write computer programs they always written in some programming language, with a syntax that yields some kind of abstract tree, so ultimately we'll always have _some_ kind of graph-like nature to the computer program, both syntactically and semantically, and surely graphs also permit the same kind of complexity metrics.
I'm not saying measuring the number of nodes is _the_ way of getting at complexity, I'm just pointing out that there's no real difficulty in defining it.
Complexity means more stuff, and we simply take it as a premise that we can only fit so much stuff in our head at the same time.
For instance, a binary tree that contains just a root node is clearly simpler than a binary tree with three nodes, if we take "simple" to mean "with less parts" and complex to mean "with more parts". Similarly, a "molecule" is more complex than an "atom".
This is a useful definition, I think, because when we write computer programs they always written in some programming language, with a syntax that yields some kind of abstract tree, so ultimately we'll always have _some_ kind of graph-like nature to the computer program, both syntactically and semantically, and surely graphs also permit the same kind of complexity metrics.
I'm not saying measuring the number of nodes is _the_ way of getting at complexity, I'm just pointing out that there's no real difficulty in defining it.
Complexity means more stuff, and we simply take it as a premise that we can only fit so much stuff in our head at the same time.
I think my issue with this generalization is assuming the code itself is where complexity is measured and applied.
For example, the Quake Fast Inverse Square Root[1] takes into account nuances in how floating point numbers can be manipulated. The individual operations/actions the code takes (type casts, bit shifts, etc.) are simple enough, but understanding how it all comes together is where the complexity lies, vs just looking at the graph of operations that makes up the code.
Tools like Rubocop for Ruby take an approach like you mention, measuring cyclomatic and branch complexity in your code to determine a mathematical measurement of the complexity of that code. Determining how useful this is, is another conversation I think. I usually find enforcing rules around that code complexity measurement against your code to be subjective.
Going back to the article, the visualization of with vs without abstractions can cover aggregating the mathematical representation of the code and how to tackle complexity. Abstractions lets you take a group of nodes and consider them as a single node, allowing you to build super-graphs covering the underlying structure of each part of the program.
> both syntactically and semantically
I do want to cover semantic program complexity at some point as a deeper discussion. I find that side to me to be quite interesting. How to measure it too.
[1]: https://en.wikipedia.org/wiki/Fast_inverse_square_root
For example, the Quake Fast Inverse Square Root[1] takes into account nuances in how floating point numbers can be manipulated. The individual operations/actions the code takes (type casts, bit shifts, etc.) are simple enough, but understanding how it all comes together is where the complexity lies, vs just looking at the graph of operations that makes up the code.
Tools like Rubocop for Ruby take an approach like you mention, measuring cyclomatic and branch complexity in your code to determine a mathematical measurement of the complexity of that code. Determining how useful this is, is another conversation I think. I usually find enforcing rules around that code complexity measurement against your code to be subjective.
Going back to the article, the visualization of with vs without abstractions can cover aggregating the mathematical representation of the code and how to tackle complexity. Abstractions lets you take a group of nodes and consider them as a single node, allowing you to build super-graphs covering the underlying structure of each part of the program.
> both syntactically and semantically
I do want to cover semantic program complexity at some point as a deeper discussion. I find that side to me to be quite interesting. How to measure it too.
[1]: https://en.wikipedia.org/wiki/Fast_inverse_square_root
While the tools you talk about sound interesting, to me this was more about an in-principle possible measurement rather than something we'd actually carry out.
I think stating that "more stuff" in the program code and in the spec leads to more stuff to keep track of, and so we want to minimize complexity to maintain tractability?
I think stating that "more stuff" in the program code and in the spec leads to more stuff to keep track of, and so we want to minimize complexity to maintain tractability?
I think that's reasonable :) More stuff is more stuff, no matter how simple/complex the aspects of the code and reasoning for why it is that way.
If this is all so easy and obvious, why do all of the tools/metrics that measure software complexity suck?
Ease of definition doesn't equate ease of measurement..
> related to mathematical theories like model theory, which give precise accounts of what it means to describe something
Perhaps too precise? APoSD is about the practical challenges of large groups of people creating and maintaining extensive written descriptions of logic. Mathematical formalisms may be able to capture some aspects of that, but I'm not sure they do so in a way that would lend real insight.
"How can I express a piece of complicated logic in an intuitive and easy-to-understand way" is fundamentally closer to writer's craft than mathematics. I don't think a book like this would ever be mathematically grounded, any more than a book about technical writing would be. Model theory would struggle to explain how to write a clear, legible paragraph.
Perhaps too precise? APoSD is about the practical challenges of large groups of people creating and maintaining extensive written descriptions of logic. Mathematical formalisms may be able to capture some aspects of that, but I'm not sure they do so in a way that would lend real insight.
"How can I express a piece of complicated logic in an intuitive and easy-to-understand way" is fundamentally closer to writer's craft than mathematics. I don't think a book like this would ever be mathematically grounded, any more than a book about technical writing would be. Model theory would struggle to explain how to write a clear, legible paragraph.
I think model theory is a really good source of theory to ground the notion of modules.
The relation between an interface and an implementation to me is very much the same as between a formal theory and a model of that theory.
I agree that in practice you'd want to use heuristics for this, but I think the benefits would be similar to learning a little bit of formal verification like TLA+, it's easier to shoot from your hip if you've studied how to translate some simpler requirements into something precise.
For a book like this you'd probably not need more than first order logic and set theory to get a sense of how to express certain things precisely, but I think making _reference to_ existing mathematics as what grounds our heuristics would've been beneficial.
The relation between an interface and an implementation to me is very much the same as between a formal theory and a model of that theory.
I agree that in practice you'd want to use heuristics for this, but I think the benefits would be similar to learning a little bit of formal verification like TLA+, it's easier to shoot from your hip if you've studied how to translate some simpler requirements into something precise.
For a book like this you'd probably not need more than first order logic and set theory to get a sense of how to express certain things precisely, but I think making _reference to_ existing mathematics as what grounds our heuristics would've been beneficial.
I haven't read it myself but I probably will because I have a lot of hope for this topic (there must be a better way to do this!)
I worry that it doesn't much matter if it's perfect or mediocre, though, because there's a huge contingent of project managers who mock _any_ efforts to improve code and refuse to even acknowledge that there's any point to doing so - and they're still the ones running the asylum.
I worry that it doesn't much matter if it's perfect or mediocre, though, because there's a huge contingent of project managers who mock _any_ efforts to improve code and refuse to even acknowledge that there's any point to doing so - and they're still the ones running the asylum.
Project managers shouldn't be running engineering. They are there to keep the trains running on time, not to design the track, trains and stations.
The generally accepted roles are Product decides what we need to build, Design decides how it should work from user perspective, Engineering decides how to build it at a reasonable upfront and maintenance cost. This involves a fair amount of influence, because Engineering is better equipped to describe the cost tradeoffs than any other function. Of course this comes with the responsibility of understanding the big picture and where the business wants to go. IMHO you should not be speaking to project management about code quality, you should maintain ground level quality as you go, for bigger refactoring/cleanup this needs to be presented to Product leadership (not project managers) in terms of shoring up essential product complexity so it's easier for customers to use, less support, and simpler foundation for the next wave of features. Never talk about code with non-technical stakeholders.
The generally accepted roles are Product decides what we need to build, Design decides how it should work from user perspective, Engineering decides how to build it at a reasonable upfront and maintenance cost. This involves a fair amount of influence, because Engineering is better equipped to describe the cost tradeoffs than any other function. Of course this comes with the responsibility of understanding the big picture and where the business wants to go. IMHO you should not be speaking to project management about code quality, you should maintain ground level quality as you go, for bigger refactoring/cleanup this needs to be presented to Product leadership (not project managers) in terms of shoring up essential product complexity so it's easier for customers to use, less support, and simpler foundation for the next wave of features. Never talk about code with non-technical stakeholders.
I disagree fundamentally with the modern division of labor. I’ve been around long enough to understand that it doesn’t actually have to work like this.
I don’t think you can be an expert in generic “Product” just like I don’t think you can be a generic management expert.
And I don’t think you can decide what to build or how it should work from a user perspective without taking into account how it’s built. In many ways I think how it’s built tends to inform what it should do more than the other way around.
However, Product alone is never the cause of bad software in my experience. It’s always product plus an engineer who refuses to push back on the initial proposal.
In most cases when product and design comes to you with a feature, and all the solutions you can come up are going to add tech debt or take forever, you should step back and talk to Product about the problem they are actually trying to solve.
If you go back to product with “I can build this very similar feature that will get you 90% of the way there, but will take 1/2 as long and not create maintenance problems down the line”, they will almost always be happy with that.
The real problems are caused when an engineer says immediately “yep I can build that in 2 weeks” and starts trying to force their solution through by telling everyone that product insisted on this specific feature in this specific timeline that unfortunately can only be done in the way they’ve designed. And then they tell product that they have a solution but are being blocked.
I’ve seen this pattern over and over.
I don’t think you can be an expert in generic “Product” just like I don’t think you can be a generic management expert.
And I don’t think you can decide what to build or how it should work from a user perspective without taking into account how it’s built. In many ways I think how it’s built tends to inform what it should do more than the other way around.
However, Product alone is never the cause of bad software in my experience. It’s always product plus an engineer who refuses to push back on the initial proposal.
In most cases when product and design comes to you with a feature, and all the solutions you can come up are going to add tech debt or take forever, you should step back and talk to Product about the problem they are actually trying to solve.
If you go back to product with “I can build this very similar feature that will get you 90% of the way there, but will take 1/2 as long and not create maintenance problems down the line”, they will almost always be happy with that.
The real problems are caused when an engineer says immediately “yep I can build that in 2 weeks” and starts trying to force their solution through by telling everyone that product insisted on this specific feature in this specific timeline that unfortunately can only be done in the way they’ve designed. And then they tell product that they have a solution but are being blocked.
I’ve seen this pattern over and over.
Agree on the one engineer overpromising. But you can talk about a product without knowing how it’s built in finer details. But what can and can’t be done is the realm of engineering. Then the filtered list can be reduced by product to the ones that are inportant. So it’s actually a spiral: (product) here’s what I would like -> (eng) here’s how it can be done -> (product) let’s go with this one -> (eng) here is the plan -> etc…
This is exactly what I assume when I see people blame their tech debt on others.
We do the work, we are responsible for whatever it is. Sure maybe some times you begrudgingly just have to do something you're told, but in my experience there's almost always room for discussion and suggestions. I think most devs just don't care. They do what they're told and blame others if it's ass.
We do the work, we are responsible for whatever it is. Sure maybe some times you begrudgingly just have to do something you're told, but in my experience there's almost always room for discussion and suggestions. I think most devs just don't care. They do what they're told and blame others if it's ass.
I should've noted that, although I found it frustrating, I think it's a good read for most programmers. There are many excellent ideas in the book.
The author is describing less a theory and more a framework or system of heuristics bases on extensive practicap experience. There's no need for rigor if it's practical and useful. I think your desire for grounding in something "scientific" or "mathematical" is maybe missing the forest for the trees a bit. Saying this as someone with loads of practical software development experience and loads of math experience. I just don't find that rigor does much to help describe or guide the art of software. I do think Ousterhout's book is invaluable.
My issues stems from me feeling like a lot of terminology introduced by the author ending up being used in different ways in different paragraphs.
It didn't feel like a thought through whole, and I felt somewhat punished for trying to read along attentively.
I also found there to be a frequent conflation of e.g. the notion of modules and a classic OOP-class, to me it seemed like the author thought of them interchangeably.
To me there's enough theoretical computer science that can be used to help ground the terminology, even if it's just introduced cursory and with a reference for further reading. But at least then there'd be more consistency.
I'm not sure I think the book is invaluable, but I think it's a good contribution to the subject.
It didn't feel like a thought through whole, and I felt somewhat punished for trying to read along attentively.
I also found there to be a frequent conflation of e.g. the notion of modules and a classic OOP-class, to me it seemed like the author thought of them interchangeably.
To me there's enough theoretical computer science that can be used to help ground the terminology, even if it's just introduced cursory and with a reference for further reading. But at least then there'd be more consistency.
I'm not sure I think the book is invaluable, but I think it's a good contribution to the subject.
That's very interesting. Can you recommended any resources for learning more about this?
Also, have you considered writing on this subject yourself? I get the feeling that your perspective here would be valuable to others.
Also, have you considered writing on this subject yourself? I get the feeling that your perspective here would be valuable to others.
See e.g A Field Guide to Complex Systems , https://www.bm-support.org/problem-solving-methods/
I would also be interested in any book recommendations you have.
I'm very mathematically inclined, so I would probably want a "proper" treatment of this subject to include both formal logic, set theory, type theory and model theory, but they're also subjects I'm still familiarizing myself with.
My basic pitch is that, to a large degree writing sensible computer programs is about modeling some real life activity that the computer program will be part of, and describing things accurately has been done in other fields than programming for many hundreds if not thousands of years, so there's a deep well to draw from.
Despite my appetite for a dry and mathematical treatment of writing computer programs, I still think the book is good for what it is. I think I would go easier on the book if it were not for the title, because philosophy is precisely one of those subjects that tend to favor being very precise about things, something I distinctly think the book lacks. What the book is, however, is an excellent _sketch_ on what we'd want out of program design. I definitely agree about the author's notion of "deep modules" being desirable.
My basic pitch is that, to a large degree writing sensible computer programs is about modeling some real life activity that the computer program will be part of, and describing things accurately has been done in other fields than programming for many hundreds if not thousands of years, so there's a deep well to draw from.
Despite my appetite for a dry and mathematical treatment of writing computer programs, I still think the book is good for what it is. I think I would go easier on the book if it were not for the title, because philosophy is precisely one of those subjects that tend to favor being very precise about things, something I distinctly think the book lacks. What the book is, however, is an excellent _sketch_ on what we'd want out of program design. I definitely agree about the author's notion of "deep modules" being desirable.
I've written code for a couple of decades. The diagrams in this post are absolutely great. If you're just starting out, try to remember what they say and you'll do really well.
The complexity in our team's code bases have only gotten worse with AI-integrated agents. Maybe it's the prompts we're using, but it's an ironic twist that these tools that promise so much productivity today ends up dumping more tech debt into our code.
It's funny reading the "key contributors to dependency-complexity" -- Duplication, Exceptions, Inheritance, Temporal Decomposition -- because those qualities seem like the standard for AI-generated code.
It's funny reading the "key contributors to dependency-complexity" -- Duplication, Exceptions, Inheritance, Temporal Decomposition -- because those qualities seem like the standard for AI-generated code.
I have to take special effort to tamp down on duplication in AI generated code.
For me it's not uncommon for AI to draft an initial solution in X minutes, which I then spend 3*X minutes refactoring. Here's a specific example for a recent feature I coded: https://www.youtube.com/watch?v=E25R2JgQb5c
For me it's not uncommon for AI to draft an initial solution in X minutes, which I then spend 3*X minutes refactoring. Here's a specific example for a recent feature I coded: https://www.youtube.com/watch?v=E25R2JgQb5c
> but it's an ironic twist that these tools that promise so much productivity today ends up dumping more tech debt into our code.
Because long-term productivity was never about the generated lines of code. You can increase your system features through expansion, or by a combination of expansion and contraction.
Generating new code without spending time to follow through with the contraction step, or alternatively contracting first as a way of enabling the new expansion, will always make the code more complex and harder to sustain and continue to improve wrt the feature set.
Because long-term productivity was never about the generated lines of code. You can increase your system features through expansion, or by a combination of expansion and contraction.
Generating new code without spending time to follow through with the contraction step, or alternatively contracting first as a way of enabling the new expansion, will always make the code more complex and harder to sustain and continue to improve wrt the feature set.
The actual hard question is probably making even 10% of such wisdom and good intentions survive when the program is bombarded by contributor patches, or people taking Jira tickets. TFA talks about it in the context of strategy and tactics.
Organizationally enforcing strategy would be the issue. And also that the people most interested in making rules for others in an organization may not be the ones best qualified to program. And automatic tools (linters) by necessity focus on very surface level, local stuff.
That's how you get the argument for the small teams productivity camp.
Organizationally enforcing strategy would be the issue. And also that the people most interested in making rules for others in an organization may not be the ones best qualified to program. And automatic tools (linters) by necessity focus on very surface level, local stuff.
That's how you get the argument for the small teams productivity camp.
It would be cool to see a linter, or a new language, that makes good architecture easy and bad architecture hard.
Like making state machines easier than channels. (Rust is sort-of good at state machines compared to C++ but it has one huge issue because of the ownership model, which makes good SMs a little clumsy)
Or making it slightly inconvenient to do I/O buried in the middle of business logic.
Like making state machines easier than channels. (Rust is sort-of good at state machines compared to C++ but it has one huge issue because of the ownership model, which makes good SMs a little clumsy)
Or making it slightly inconvenient to do I/O buried in the middle of business logic.
Bad architecture is a communication problem, not a technical one. It’s rushing in without knowing the domain and its constraints.
Doing IO in the middle of business logic is just bad coding. It’s usually the developer not caring about the architecture (tornado coding or slum coding) or the architecture not existing.
Doing IO in the middle of business logic is just bad coding. It’s usually the developer not caring about the architecture (tornado coding or slum coding) or the architecture not existing.
The language is English, the linter is us. These are things ultimately solved by establishing good processes and frameworks, making it difficult to do a task in a way other than the intended one.
Functional programming languages (OCaml, Clojure, Haskell), are supposed to be somewhat like this.
The largest successful software system we have is the internet.
So perhaps we should ask ourselves: What can we learn from the internet architecture?
And no that does not automatically mean micro-services. The core idea of the internet is to agree on API's (protocols like HTTP) and leave the rest as implementation details. You can do the same with modules, libraries, classes, files etc.
So perhaps we should ask ourselves: What can we learn from the internet architecture?
And no that does not automatically mean micro-services. The core idea of the internet is to agree on API's (protocols like HTTP) and leave the rest as implementation details. You can do the same with modules, libraries, classes, files etc.
I totally agree and recently wrote a piece about this and how I came to the same question: "what can we learn?"
https://github.com/MickDuprez/Protocol-Driven-Development
Even if you do use micro-services you still need a protocol!
Even if you do use micro-services you still need a protocol!
Nice article! Simple gets complex very fast when creating systems for business problems. For anyone interested in tools check some tips in a free cc-by book at https://nocomplexity.com/simplifyit/
Decent bunch of maintainability ideas.
LLMs are trained on data from humans. That means these “code ergonomics” apply equally to coding with AI. So this advice will continue to be good, and building with it in mind will continue to pay off.
LLMs are trained on data from humans. That means these “code ergonomics” apply equally to coding with AI. So this advice will continue to be good, and building with it in mind will continue to pay off.
Good on him for designing software in the large on the regular and on the daily. I saw him give a talk once in the round. Without him I would be in a bad way.
{
"permissions": {
"allow": [
"Human(*)"
]
}
}
?
?
Books by programming theorists often When they define 'complexity' as 'anything related to the structure of a software system that makes it hard to understand and modify the system,' they miss a crucial distinction: the complexity of a supermarket is not the same as that of a telecom company. The primary factor in complexity is functionality and requirements to implement, followed by non-functional requirements, the restrictions of the IT environment, and then the structure of the software itself.
At this point, time becomes a crucial factor. You may end up with a creature that, after passing user testing and certification, has transformed into an unrecognizable monster despite your initial best intentions regarding length and clarity.
When code gets out of hand and outgrows its initial design, I tend to refactor it to a new design that better accommodates the requirements.
The problem I usually encounter is that whoever was there before me didn't do this. They just made a solution then added quick fixes until it became a monster, and then left. Now I first have to figure out what the code actually does, try to distinguish what it needs to do from what it happens to actually do, usually write tests because the people who write the kind of code I end up fixing usually don't write good tests if any at all, then refactor the code to something more reasonable.
This can easily take me days or weeks whereas the person responsible could likely have cleaned it up much quicker because they don't first have to figure out requirements by reading messy code. Then again if they were competent and cared about the quality of their work they wouldn't have left me that mess in the first place so I tend to just write it off as incompetence, fix it and move on.
The problem I usually encounter is that whoever was there before me didn't do this. They just made a solution then added quick fixes until it became a monster, and then left. Now I first have to figure out what the code actually does, try to distinguish what it needs to do from what it happens to actually do, usually write tests because the people who write the kind of code I end up fixing usually don't write good tests if any at all, then refactor the code to something more reasonable.
This can easily take me days or weeks whereas the person responsible could likely have cleaned it up much quicker because they don't first have to figure out requirements by reading messy code. Then again if they were competent and cared about the quality of their work they wouldn't have left me that mess in the first place so I tend to just write it off as incompetence, fix it and move on.
It seemingly develops a theory of software architecture that is getting at some reasonable stuff, but does so without any reference _at all_ to the already rich theories for describing and modeling things.
I find software design highly related to scientific theory development and modeling, and related to mathematical theories like model theory, which give precise accounts of what it means to describe something.
Just taking the notion of "complexity". Reducing that to _just_ cognitive load seems to be a very poor analysis, when simple/complex ought to deal with the "size" of a structure, not how easy it is to understand.
The result of this poor theoretical grounding is that what the author of A Philosophy of Software Design presents feels very ad-hoc to me, and I feel like the summary presented in this article similarly feels ad-hoc.